HTTP API

• POST /v1/chat/completions — OpenAI-compatible, streaming (SSE) included.
• GET /v1/models — the full catalog (models, agent variants, queues), with owned_by providers.
• POST /v1/workers/register · GET /v1/workers/{name}/jobs?wait= · POST /v1/workers/{name}/jobs/{id} · GET /v1/workers — the pull-worker surface.
• GET /health — always open.

Streaming

routsi supports both a plain sync response and "stream": true (standard OpenAI SSE chunks). On a streaming request it also sends SSE comment lines — : ping — roughly every 15 seconds (configurable via stream_heartbeat, 0 disables them), so a long agent or worker turn doesn't get dropped by an idle proxy/load-balancer timeout.

data: {"choices":[{"delta":{"content":"thinking"}}]}

: ping

: ping

data: {"choices":[{"delta":{"content":" ...done"}}]}

data: [DONE]

Because : ping lines are SSE comments (not data: events), the OpenAI SDK's stream parser silently skips them — no client-side changes needed.

Next: Scenarios →